{
  "lesson": "29-moderation-systems-openai-perspective-llamaguard",
  "title": "Moderation Systems - OpenAI, Perspective, Llama Guard",
  "questions": [
    {
      "stage": "pre",
      "question": "What is OpenAI's omni-moderation-latest (2024)?",
      "options": [
        "A C2PA watermarking endpoint",
        "A GPT-4o-based moderation classifier that handles text and images in one call, returning a 13-category response schema (harassment, hate, self-harm, sexual, violence, illicit, with sub-categories), free for most developers",
        "An RLHF reward model",
        "A reasoning-focused frontier model"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which of these is NOT one of OpenAI Moderation's primary top-level categories?",
      "options": [
        "election-interference",
        "violence",
        "self-harm",
        "harassment"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What are the three layers of the standard 2026 moderation pattern?",
      "options": [
        "Input moderation (pre-generation), Output moderation (post-generation), Custom moderation (domain rules)",
        "Telescopic / periscopic / microscopic",
        "Probe / detector / harness",
        "Tokenizer / model / detokenizer"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What is Perspective API (Google Jigsaw) primarily designed to score?",
      "options": [
        "Reward over user satisfaction",
        "Reasoning quality",
        "Watermark strength",
        "Toxicity (with severe-toxicity, insult, profanity, threat, identity-attack sub-dimensions) as a pre-LLM-era content-moderation baseline"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What is one failure mode of input-only moderation?",
      "options": [
        "It does not catch model output failures or hallucinations; encoding attacks (Lessons 12-14) can bypass input classifiers entirely",
        "It cannot run on GPUs",
        "It always blocks legitimate content",
        "It removes the need for output classifiers"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What is the Azure Content Moderator deprecation and migration target?",
      "options": [
        "Merged into Llama Guard 4",
        "Replaced by Perspective API",
        "Deprecated February 2024, retired March 15, 2027; migration target is Azure AI Content Safety (LLM-based, integrated with Azure OpenAI)",
        "Deprecated immediately with no replacement"
      ],
      "correct": 2,
      "explanation": ""
    }
  ]
}